v-server change needs to trigger evaluation and persisiting of load stats #79

New issue

Open

opened 2020-10-06 14:14:20 +02:00 by schmittlauch · 0 comments

schmittlauch commented

2020-10-06 14:14:20 +02:00

Owner

problem szenario

In the middle of a load-eval wait period, a VS is moved form slice A to slice B.
That means slice A recieves load for ~= period/2 and slice B receives load for ~= period/2, but at load evaluation their load is divided by period, effectively halving the recorded load of the tag. That is especially important for reporting the load of the newly acquired tag, as that starts to increas just after joining anyways.

solution idea

Once a new VS is joined, trigger a new load evaluation to finish the period in which vs B was not active and start a period in which vs B is active.

potential problems

if this happens too quickly after the last load evaluation, the period length is unrealistically small, giving unrealistic load results
maybe the whole method of load evaluation at discrete points is flawed?
maintain a list of previous loads and do a weighted average over them - that'd raise exactly this issue again, but attenuated by the weighted averaging

idea: also record load collection period in the load stats, use it as weight for averaging

## problem szenario In the middle of a load-eval wait period, a VS is moved form slice A to slice B. That means slice A recieves load for ~= period/2 and slice B receives load for ~= period/2, but at load evaluation their load is divided by period, effectively halving the recorded load of the tag. That is especially important for reporting the load of the newly acquired tag, as that starts to increas just after joining anyways. ## solution idea Once a new VS is joined, trigger a new load evaluation to finish the period in which vs B was not active and start a period in which vs B is active. ### potential problems - if this happens too quickly after the last load evaluation, the period length is unrealistically small, giving unrealistic load results - maybe the whole method of load evaluation at discrete points is flawed? - maintain a list of previous loads and do a weighted average over them - that'd raise exactly this issue again, but attenuated by the weighted averaging idea: also record load collection period in the load stats, use it as weight for averaging